Remaining useful life prediction for solid-state lithium batteries based on spatial–temporal relations and neuronal ODE-assisted KAN

Remaining useful life prediction (RUL) of solid-state lithium batteries (SSLIBs) can accelerate the maintenance and optimization process, facing challenges in insufficient exploration of implicit degradation information, complexity of computational costs and poor interpretability. To address these issues, a novel method for obtaining comprehensive implicit information during the degradation process is proposed. Firstly, topological relations are introduced by using graph attention network (GAT) to comprehensively represent the implicit relations among external parameters. It is utilized to supplement the interdependencies between physical measurements of multiple health indicators for SSLIBs, avoiding manual feature engineering. Then, a neural ordinary differential equation (ODE) composed of Kolmogorov–Arnold network (KAN) is developed to capture the continuous dynamic implicit state trajectories during the degradation process, overcoming the issue of ignoring dynamic variations for implicit relations in external parameters. Moreover, KAN is adopt as a regressor, which ensures the interpretability of the constructed RUL prediction model for SSLIBs while reducing the computational cost. The comparison analysis in the real SSLIBs degradation datasets demonstrate the optimal minimum root mean square errors and the parameters of the model are reduced by 39.03% and 49.13%, respectively. It also indicates that the proposed method can provide new perspectives and solutions for RUL prediction of SSLIBs.